Most fraud models are trained on a moment — a transaction, an application. Wakeful's core models are trained on trajectories: how an identity's data footprint evolves over months, and what separates organic thinness from engineered dormancy.
The difference only appears in the shape of the trajectory: a real consumer's data footprint grows organically and idiosyncratically. A cultivated identity's footprint grows on a schedule — because it's following a playbook, even an adaptive one.
A sequence model trained on longitudinal identity behavior — credit utilization cadence, address change timing, device continuity — that scores deviation from organic consumer trajectories at every point in an identity's life, not just at origination.
A heterogeneous graph neural network over devices, addresses, application metadata, and (where consortium data is available) cross-institution signals, trained to surface clusters that share a manufacturing fingerprint rather than a coincidental overlap.
A multimodal model combining document image forensics, active liveness challenge-response, and generative-artifact detection tuned specifically to AI-generated identity documents and deepfake selfie injection attacks — the fastest-growing onboarding attack vector.
Off-the-shelf fraud stacks are tuned to score a transaction or an application in isolation, because that's where most fraud losses show up. Synthetic identity fraud is a multi-year cultivation problem hiding inside a system built to reason in milliseconds. Solving it requires infrastructure that can hold, index, and continuously re-score an entire portfolio's history — not bolt a longer lookback window onto a transaction-scoring engine.
Every production model ships with SHAP-based feature attribution, a documented training lineage, drift monitoring against your live portfolio, and a challenger-model process — the documentation package your model risk management function needs for SR 11-7 and OCC 2011-12 review, generated automatically rather than reconstructed after the fact.
Wakeful deploys as a private, single-tenant instance inside your cloud environment or ours, under your data residency and retention policy. Raw PII never leaves your perimeter unless you explicitly enable consortium matching.
Every identity in your book is re-evaluated on a rolling basis, not just at origination — so a shift from dormant to activating is caught the week it starts, not the quarter it becomes a loss.
Institutions that opt into privacy-preserving consortium matching — hashed device and identity-element overlap, never raw PII — sharpen cluster detection for every participant without exposing customer data across institutional boundaries.